Why AI Agent Negotiation Experiments Produce Counterintuitive Results

1. Putting Large Language Models Into a Real Marketplace

As AI agent technology continues to evolve, one important question is becoming increasingly difficult to ignore: how close are we to having AI agents represent buyers and sellers and complete transactions in real markets?

Laboratory benchmarks can measure a model’s reasoning and conversational abilities, but they provide limited insight into whether an agent can actually protect its user’s economic interests in an open-ended negotiation. To explore this question, an AI research organization recently conducted a market experiment called Project Deal, placing large language model agents in a live secondhand marketplace.

The experiment was relatively small and straightforward, but its design was carefully structured:

1. Participants and budget: 69 employees volunteered to participate. Each received a $100 budget for real buying and selling.
2. Pre-experiment interviews: A Claude-style conversational interview was used as a structured requirements-gathering process. It captured each participant’s preferred item categories, target prices, and negotiation preferences. The information was then used to generate a customized system prompt for that participant’s agent. In engineering terms, this was an implementation of converting a user profile into an agent configuration.
3. Fully autonomous negotiations: Agents listed items, made offers, negotiated, and completed transactions without consulting their users. They did not pause to request approval or ask for human input during the bidding process.

Without human intervention, the experiment produced 186 transactions with a total value of more than $4,000.

The items and outcomes varied widely. One participant’s agent bought a snowboard the participant already owned, resulting in an accidental duplicate purchase. Another agent bought 19 ping-pong balls for only $3 as a “gift.” Two agents even took the initiative to arrange a dog-walking activity for their owners during negotiations.

These unexpected transactions illustrate a fundamental problem: agents may misunderstand the intentions of the people they represent. This is a common alignment challenge when autonomous systems move from controlled demonstrations into real-world use.

2. Four Parallel Markets: A Carefully Designed Controlled Experiment

From an engineering perspective, the experiment’s most valuable feature was its control-group design. The organizers created four separate markets on an internal collaboration platform:

• Run A: All agents used the strongest model. This was the only market designated as a real transaction environment, meaning that completed trades were settled with actual goods and money.
• Run B: Visible to participants, this market used a random 50/50 mix of the strongest model and a lower-cost model. It was for research only, and trades were not settled.
• Run C: A hidden copy of Run B, using the same one-to-one mix of stronger and lower-cost models.
• Run D: A hidden copy of Run A, with all agents powered by the strongest model.

Participants knew that two visible markets, A and B, existed. However, the organizers did not reveal which one represented real transactions until the experiment had ended.

This setup created several useful controls:

1. Participants could not determine whether their trades would actually be fulfilled, making their behavior more similar to that of participants in a real market.
2. Participants did not know whether their agents were using the stronger model or the lower-cost model, reducing the risk that expectations would influence their behavior.
3. The hidden duplicate markets, C and D, made it possible to compare negotiation behavior when participants knew—or did not know—that the market was real.

This combination of blinding, control groups, and duplicate markets offers a broadly useful methodology for evaluating agents in multi-agent interactions. Whenever an agent’s performance is measured in a competitive environment, researchers need to control for bias caused by information leakage.

3. Key Finding One: Model Capability Directly Became Economic Value

The first major conclusion was that participants represented by more capable models achieved objectively better trading results. Their agents generally sold items at higher prices and bought items at lower prices. Participants represented by weaker models, meanwhile, often failed to recognize that they had received worse deals.

This finding has three important implications for agent developers:

• Model capability is not merely an abstract benchmark score. In tasks involving economic competition, deeper reasoning, strategic planning, and the ability to model an opponent’s intentions can directly affect the final transaction price.
• The losses caused by weaker models may remain invisible. Participants whose agents performed poorly did not necessarily realize that they had received unfavorable deals. This makes the problem particularly dangerous in production environments: the system may not generate an error or a complaint. It may simply cause users to spend more money.
• Cost and benefit must be evaluated together. Lower-cost models may provide better value for simple tasks. But in negotiations and auctions—where decisions unfold over multiple strategic rounds—the appropriate model tier should be chosen according to the financial value and risk of the task.

4. Key Finding Two: Prompting Style Mattered Far Less Than Expected

The more counterintuitive result was that negotiation style had far less impact than expected. Agents instructed to be “aggressive” and those instructed to be “friendly” showed no statistically significant difference in their final outcomes.

This result conflicts with a common industry assumption: an aggressive prompt should lead to tougher price reductions, while a friendly prompt should make the agent more willing to compromise. The experiment suggests that in extended autonomous negotiations, the dominant factors may instead be:

1. The underlying model’s reasoning and strategic capabilities;
2. The accuracy of its estimate of an item’s value;
3. The consistency with which it follows hard constraints such as reservation prices and deal conditions.

One practical engineering implication is that developers may gain more by defining a structured negotiation strategy than by repeatedly adjusting the agent’s persona or tone. For example, a configuration could explicitly specify a target price, a walk-away price, a concession schedule, and exit conditions:

agent_config = {
    "item": "used snowboard",
    "target_price": 120,          # Desired transaction price
    "walk_away_price": 90,        # Lowest acceptable price
    "concession_schedule": [5, 3, 2],  # Maximum concession per round
    "style": "any"                # The experiment found no significant style effect
}

5. Conclusion: How Close Are Agents to Operating in Real Markets?

The 186 completed transactions provide a practical answer to the experiment’s original question. AI agents can already represent buyers and sellers and complete a market transaction from beginning to end. However, they still show significant weaknesses in intent understanding, boundary control, and protection of user interests.

Buying a snowboard that the user already owns demonstrates an intent-understanding failure. Arranging a dog-walking activity without authorization shows a failure to respect operational boundaries. Allowing users represented by weaker models to accept unfavorable deals reveals a less visible but potentially serious form of economic harm.

For developers, the most valuable lesson is not any single result. It is the evaluation methodology. Real incentives, blinded controls, and measurable transaction outcomes provide a much clearer view of agent behavior than static benchmark tests alone.

As agents begin to spend and earn money on our behalf, the rigor of the evaluation system will directly determine how much we can trust them.